Get rid of the superstitious "~" in dict hashing's "i = (~hash) & mask".

The comment following used to say:
	/* We use ~hash instead of hash, as degenerate hash functions, such
	   as for ints <sigh>, can have lots of leading zeros. It's not
	   really a performance risk, but better safe than sorry.
	   12-Dec-00 tim:  so ~hash produces lots of leading ones instead --
	   what's the gain? */
That is, there was never a good reason for doing it.  And to the contrary,
as explained on Python-Dev last December, it tended to make the *sum*
(i + incr) & mask (which is the first table index examined in case of
collison) the same "too often" across distinct hashes.

Changing to the simpler "i = hash & mask" reduced the number of string-dict
collisions (== # number of times we go around the lookup for-loop) from about
6 million to 5 million during a full run of the test suite (these are
approximate because the test suite does some random stuff from run to run).
The number of collisions in non-string dicts also decreased, but not as
dramatically.

Note that this may, for a given dict, change the order (wrt previous
releases) of entries exposed by .keys(), .values() and .items().  A number
of std tests suffered bogus failures as a result.  For dicts keyed by
small ints, or (less so) by characters, the order is much more likely to be
in increasing order of key now; e.g.,

>>> d = {}
>>> for i in range(10):
...    d[i] = i
...
>>> d
{0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 5, 6: 6, 7: 7, 8: 8, 9: 9}
>>>

Unfortunately. people may latch on to that in small examples and draw a
bogus conclusion.

test_support.py
    Moved test_extcall's sortdict() into test_support, made it stronger,
    and imported sortdict into other std tests that needed it.
test_unicode.py
    Excluced cp875 from the "roundtrip over range(128)" test, because
    cp875 doesn't have a well-defined inverse for unicode("?", "cp875").
    See Python-Dev for excruciating details.
Cookie.py
    Chaged various output functions to sort dicts before building
    strings from them.
test_extcall
    Fiddled the expected-result file.  This remains sensitive to native
    dict ordering, because, e.g., if there are multiple errors in a
    keyword-arg dict (and test_extcall sets up many cases like that), the
    specific error Python complains about first depends on native dict
    ordering.

1 files changed, 10 insertions, 0 deletions

diff --git a/Misc/NEWS b/Misc/NEWS
index b4859599e4..3491109b34 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -23,6 +23,16 @@ Core
   usually for the better, but may also cause numerically unstable
   algorithms to break.
 
+- The implementation of dicts suffers fewer collisions, which has speed
+  benefits.  However, the order in which dict entries appear in dict.keys(),
+  dict.values() and dict.items() may differ from previous releases for a
+  given dict.  Nothing is defined about this order, so no program should
+  rely on it.  Nevertheless, it's easy to write test cases that rely on the
+  order by accident, typically because of printing the str() or repr() of a
+  dict to an "expected results" file.  See Lib/test/test_support.py's new
+  sortdict(dict) function for a simple way to display a dict in sorted
+  order.
+
 - Dictionary objects now support the "in" operator: "x in dict" means
   the same as dict.has_key(x).